Recurrence, routing, compression, and specialized hardware target the next wave of inference gains

Recent work spans loop-native recurrent reasoning, token-level routing, sparse-first serving, autoregressive-calibration pruning, trajectory distillation, vision-token compression, and photonic inference. Reported gains include more than 20,000 useful recurrent steps, 2.01x–64.15x routing throughput improvements, and strong low-parameter optical classification, but hardware dependence, workload sensitivity, and end-to-end energy claims require independent reproduction.